Compression loss 1 − EVR(r) of the training data (normalized HXS’s of one assembly) in function of the truncation rank r of the SVD, where EVR(r) is its cumulative explained variance ratio. The erratic behavior of the left part of subfigure b/ is due to the compression loss reaching single-float precision (while figure a/ was computed in double-float precision). (a) Standard-discretized dataset. (b) Finely-discretized dataset.